Agent Foundry / Custom Model Training
Build a Model You Own
Turn your best examples into a focused model your team can test, govern, and improve.
Keep the training recipe, test set, version history, and release record together.
What this page delivers
A model workshop with gates
- 1Approved examples
- 2Training run
- 3Quality gate
Result
MODEL v12 · READY
Emu on the job
Original mechanism
A model workshop with gates
Follow the work from input to a result your team can review.
Approved examples
Training run
Quality gate
Owned model
The question answered
How do we make a model fit our work without losing control?
Start with approved examples. Train a focused version. Test it against the work and rules that matter before anyone can use it.
Three-step flow
From a clear input to a governed result.
Collect the good work
Bring together approved answers, expert edits, and the rules the model must follow.
Build a focused model
Create a version for one clear job, with its recipe and source set attached.
Prove it before release
Run quality and safety checks, compare versions, then send the winner to release control.
Capabilities
The controls teams need to go deep.
Curated training sets
Keep examples clean, labeled, and tied to their source.
Version comparison
See what improved, what slipped, and what changed between runs.
Guarded training jobs
Set the data, tools, and release rules each job may use.
Portable model records
Carry the recipe, tests, approvals, and model version as one package.
Concrete product artifact
See the proof, not just a green light.
A realistic record keeps the result, the evidence behind it, and the next action in one place.
TRAINING RUN
claims-helper-v12
- Approved examples
- 8,420
- Held-out cases
- 1,200
- Policy suite
- 96 / 96
- Previous version
- +7 checks passed
› 18:42 source set locked
› 19:08 training complete
› 19:14 eval receipt signed
Buyer outcomes
Less guesswork. More control.
The recipe
Know what shaped each model version.
Every release
Only tested versions move forward.
Expert work
Turn hard-won edits into model skill.
Why Trinitite
Built for the full agent lifecycle.
Proof stays with the model
The source set, tests, and approvals travel with every version.
Safety is part of the build
Policy checks happen before release, not after a bad answer reaches a user.
Works with the wider control plane
Identity, runtime rules, evals, and release control use the same record.
Keep exploring
Connect this capability to the next handoff.
FAQ
Build a Model You Own, answered.
Do we need to train a large model from scratch?
No. Start with a focused job and use your approved examples to shape a model for that work.
Can we compare two model versions?
Yes. Each version can run against the same held-out cases and policy checks before release.
What stays with a training run?
The run record can include the source set, recipe, test results, approvals, and resulting model version.
How does a model reach production?
A passing model moves to release control, where your team sets approvals, rollout steps, and rollback rules.
Build with Agent Foundry
Build your model with proof attached.
Bring one focused use case. We will show the path from approved examples to a release-ready model.